mirror of
https://github.com/RRZE-HPC/OSACA.git
synced 2025-12-16 00:50:06 +01:00
110 lines
2.6 KiB
ArmAsm
110 lines
2.6 KiB
ArmAsm
#define INSTR vextractf128
|
|
#define NINST 64
|
|
#define N edi
|
|
#define i r8d
|
|
|
|
|
|
.intel_syntax noprefix
|
|
.globl ninst
|
|
.data
|
|
ninst:
|
|
.long NINST
|
|
.align 32
|
|
PI:
|
|
.long 0xf01b866e, 0x400921f9, 0xf01b866e, 0x400921f9, 0xf01b866e, 0x400921f9, 0xf01b866e, 0x400921f9, 0xf01b866e, 0x400921f9, 0xf01b866e, 0x400921f9, 0xf01b866e, 0x400921f9, 0xf01b866e, 0x400921f9
|
|
.text
|
|
.globl latency
|
|
.type latency, @function
|
|
.align 32
|
|
latency:
|
|
push rbp
|
|
mov rbp, rsp
|
|
xor i, i
|
|
test N, N
|
|
jle done
|
|
# create DP 1.0
|
|
vpcmpeqw xmm0, xmm0, xmm0 # all ones
|
|
vpsllq xmm0, xmm0, 54 # logical left shift: 11111110..0 (54=64-(10-1))
|
|
vpsrlq xmm0, xmm0, 2 # logical right shift: 1 bit for sign; leading mantissa bit is zero
|
|
# expand from SSE to AVX
|
|
vinsertf128 ymm0, ymm0, xmm0, 0x1
|
|
# copy DP 1.0
|
|
vmovaps ymm0, ymm0
|
|
vmovaps ymm1, ymm0
|
|
# Create DP 2.0
|
|
vaddpd ymm1, ymm1, ymm1
|
|
# Create DP 0.5
|
|
vdivpd ymm2, ymm0, ymm1
|
|
loop:
|
|
inc i
|
|
INSTR xmm3, ymm0, 1
|
|
INSTR xmm4, ymm1, 2
|
|
INSTR xmm5, ymm2, 13
|
|
INSTR xmm6, ymm0, 1
|
|
INSTR xmm7, ymm1, 2
|
|
INSTR xmm8, ymm2, 13
|
|
INSTR xmm9, ymm0, 1
|
|
INSTR xmm10, ymm1, 2
|
|
INSTR xmm11, ymm2, 13
|
|
INSTR xmm12, ymm0, 1
|
|
INSTR xmm13, ymm1, 2
|
|
INSTR xmm14, ymm2, 13
|
|
INSTR xmm15, ymm0, 1
|
|
INSTR xmm3, ymm1, 2
|
|
INSTR xmm4, ymm2, 13
|
|
INSTR xmm5, ymm0, 1
|
|
INSTR xmm6, ymm1, 2
|
|
INSTR xmm7, ymm2, 13
|
|
INSTR xmm8, ymm0, 1
|
|
INSTR xmm9, ymm1, 2
|
|
INSTR xmm10, ymm2, 13
|
|
INSTR xmm11, ymm0, 1
|
|
INSTR xmm12, ymm1, 2
|
|
INSTR xmm13, ymm2, 13
|
|
INSTR xmm14, ymm0, 1
|
|
INSTR xmm15, ymm1, 2
|
|
INSTR xmm3, ymm2, 13
|
|
INSTR xmm4, ymm0, 1
|
|
INSTR xmm5, ymm1, 2
|
|
INSTR xmm6, ymm2, 13
|
|
INSTR xmm7, ymm0, 1
|
|
INSTR xmm8, ymm1, 2
|
|
INSTR xmm9, ymm2, 13
|
|
INSTR xmm10, ymm0, 1
|
|
INSTR xmm11, ymm1, 2
|
|
INSTR xmm12, ymm2, 13
|
|
INSTR xmm13, ymm0, 1
|
|
INSTR xmm14, ymm1, 2
|
|
INSTR xmm15, ymm2, 13
|
|
INSTR xmm3, ymm0, 1
|
|
INSTR xmm4, ymm1, 2
|
|
INSTR xmm5, ymm2, 13
|
|
INSTR xmm6, ymm0, 1
|
|
INSTR xmm7, ymm1, 2
|
|
INSTR xmm8, ymm2, 13
|
|
INSTR xmm9, ymm0, 1
|
|
INSTR xmm10, ymm1, 2
|
|
INSTR xmm11, ymm2, 13
|
|
INSTR xmm12, ymm0, 1
|
|
INSTR xmm13, ymm1, 2
|
|
INSTR xmm14, ymm2, 13
|
|
INSTR xmm15, ymm0, 1
|
|
INSTR xmm3, ymm1, 2
|
|
INSTR xmm4, ymm2, 13
|
|
INSTR xmm5, ymm0, 1
|
|
INSTR xmm6, ymm1, 2
|
|
INSTR xmm7, ymm2, 13
|
|
INSTR xmm8, ymm0, 1
|
|
INSTR xmm9, ymm1, 2
|
|
INSTR xmm10, ymm2, 13
|
|
INSTR xmm11, ymm0, 1
|
|
INSTR xmm12, ymm1, 2
|
|
INSTR xmm13, ymm2, 13
|
|
INSTR xmm14, ymm0, 1
|
|
cmp i, N
|
|
jl loop
|
|
done:
|
|
mov rsp, rbp
|
|
pop rbp
|
|
ret
|
|
.size latency, .-latency |